LlmKira/fast-langdetect

Wraps Facebook's FastText models with zero external dependencies (no numpy required) and offers configurable model selection—lite for offline/memory-constrained environments (~45-60 MB) or full for higher accuracy (~170-210 MB)—with automatic fallback on memory pressure. The library handles input normalization transparently, caches loaded models per-instance for reuse, and exposes BCP-47 language codes that integrate with standard language libraries like `langcodes` or `pycountry` for localization.

300 stars. No commits in the last 6 months.